Boston University
Statistical learning methods for multi-omics data integration and analysis
Abstract
dc:description.abstractHigh-throughput technologies provide important opportunities for studying complex human diseases across multiple molecular layers, making multi-omics integration an increasing trend in biomedical research. However, multi-omics integration remains challenging because omics datasets are usually high-dimensional, heterogeneous, correlated, and measured in limited sample sizes. Although many methods have been proposed to address these challenges, they often prioritize predictive accuracy while giving less attention to other aspects such as the consistency and stability of selected features. Stable feature selection is important because features that are consistently selected across replications are more likely to be reproducible and may therefore reflect reliable biological signals rather than random variation. Results based on such features can further support interpretable insights into the analyzed data. From a practical perspective, a good model should not only achieve high predictive performance, but also be interpretable, computationally efficient, and capable of selecting features that support reproducible findings. Therefore, this dissertation develops statistical learning methods for multi-view and multi-omics integration, with an emphasis on balancing predictive performance, feature selection stability, interpretability, and computational efficiency. These methods are designed not only to improve prediction, but also to identify stable and reliable features that support reproducible and interpretable findings. The dissertation consists of three main projects, each corresponding to a proposed statistical approach for multi-omics integration. These proposed methods leverage shared structures to improve prediction performance, feature selection stability, and interpretation. The shared structure is captured through sample-level similarity or through the joint modeling of correlated outcomes. Project 1 focuses on incorporating sample-level similarity, whereas Projects 2 and 3 focus on jointly modeling homogeneous and correlated outcomes. The proposed methods from all three projects were evaluated through simulation studies and real data applications using data from a type 2 diabetes study. Overall, results from both the simulation studies and real data analyses showed that the proposed approaches achieved comparable or improved predictive performance while enhancing feature selection stability. This suggests that the proposed methods can identify more reproducible features without compromising prediction accuracy. In addition to leveraging shared structures, Project 1 examined the role of incorporating non-omics data, while Projects 2 and 3 evaluated different integration strategies, including single-omics multi-outcome, multi-omics single-outcome, and multi-omics multi-outcome approaches. In summary, this dissertation work contributes to the development of more stable and interpretable multi-omics integration frameworks, which may support more flexible analyses of complex biomedical data and bring multi-omics research closer to clinical application.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Biostatistics
- Grantor
- Boston University
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Zhang, Jingwen
- Advisor dc:contributor.advisor
-
- Liu, Ching-Ti
- ORCID dc:identifier.orcid
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53649